PydanticAI MCP Experiment

What is PydanticAI MCP Experiment?

PydanticAI MCP Experiment provides minimalist examples for running your own MCP servers with local LLM models. It uses pydantic_ai and local models via Ollama, offering four different transport configurations under the examples/ folder. This project is aimed at developers who want to get hands‑on experience with MCP clients and servers.

How to use PydanticAI MCP Experiment?

The project is managed with uv. After cloning, run uv sync to install dependencies. Then change into any example folder (e.g., examples/0_subprocess) and start the server with python server.py, followed by python client.py in another terminal. The setup uses Logfire for logging, which will prompt for authentication on first run.

Key features of PydanticAI MCP Experiment

• Four transport examples: stdio, HTTP via mcp.run, FastAPI, and Starlette
• Uses pydantic_ai and local Ollama models
• Minimal, self‑contained client/server pairs per example
• Logging integration with Logfire (optional)
• Managed with uv for reproducible environments

Use cases of PydanticAI MCP Experiment

• Experimenting with MCP servers on a local machine
• Comparing different transport protocols (stdio vs HTTP)
• Prototyping MCP‑based AI assistants with pydantic_ai
• Learning how to integrate MCP with local LLM models

FAQ from PydanticAI MCP Experiment

What is required to run the examples?

You need Python and uv. The project also assumes you have a local Ollama instance available for the LLM models.

How do I set up the project?

Run uv sync in the root directory to install all dependencies. No additional configuration is required.

How do I run an example?

Change to the example folder (e.g., examples/0_subprocess), start the server with python server.py, then in another terminal run python client.py.

What logging is used and can I change it?

The project uses Logfire by default. You will be asked to authenticate on first run. You can switch to an alternative OpenTelemetry backend like otel‑tui, but the README notes that prompt and response messages may not be displayed that way.

What MCP transport protocols are available?

The examples include stdio (via subprocess) and three HTTP variants: using mcp.run, FastAPI, and Starlette.